Development and Validation of Machine Learning-Based Prediction of Depression Progression Using EHR Data: A...

Area of Science:

• Clinical informatics
• Machine learning in healthcare
• Depression research

Background:

• Depression is a major global health issue, with timely identification of worsening cases being a significant challenge.
• Electronic health records (EHRs) offer valuable data for real-world disease trajectory analysis.
• Lack of standardized symptom scales in EHRs necessitates alternative methods like ICD10 code-based progression for predictive modeling.

Purpose of the Study:

• To develop and evaluate machine learning (ML) and deep learning (DL) models for predicting depression severity progression.
• Utilize International Classification of Diseases, 10th Revision (ICD10) codes within EHR data for predicting mild to moderate/severe depression.
• Leverage data from the MedStar Health Research Institute (MHRI) EHR database for model development and validation.

Main Methods:

• Retrospective cohort analysis of adult patients diagnosed with mild depression (ICD10) between 2017-2023.
• Inclusion of a heterogeneous feature set including demographics, socioeconomic factors, and healthcare utilization.
• Development and comparison of logistic regression, random forest, XGBoost, CatBoost, and deep neural network (DNN) models with a structured model selection framework.

Main Results:

• The analytic cohort comprised 803 patients with two-year follow-up; DNN was excluded for not meeting AUC thresholds.
• XGBoost achieved the highest composite score (Accuracy=0.72, AUC=0.776, Sensitivity=0.77), demonstrating robust predictive performance.
• Logistic regression performed closely, with other models like Random Forest being penalized for overfitting.

Conclusions:

• Machine learning models, particularly XGBoost, can effectively predict depression progression using routinely collected EHR data.
• The study highlights the feasibility of using socioeconomic and EHR data for early identification of worsening depression.
• Emphasizes the need for transparent model selection frameworks to ensure trustworthy clinical AI applications.